JP2014517390A - Support for video content search through communication network - Google Patents

Abstract

A method for supporting a user to search for video content through a communication network,
-Determining a first content item (CZ) by the user;
-Automatically extracting semantic data from the first content item;
In response to semantic data, automatically generating a query for at least one service, enabling the service or a set of videos to be retrieved by these services;
-Presenting the set content of the video to the user.

Description

  The present invention relates to access to a video content providing service through the Internet. These services are common platforms such as “Youtube”, “Videoo”, and “DailyMotion”, or more specialized platforms (education sector, medical sector, etc.).

  These two types of services share the common property of allowing users to upload their content to the platform. As a result, very large amounts of video content are available through all of these services. Currently, the “Youtube” service uploads about 355 hours of video per minute.

  For this explosive increase in available content, the vast majority of content does not gain sufficient visibility. Thus, only 40% of the available video content is viewed over 300 times.

  This low percentage is mainly due to the fact that it is very difficult to identify interesting videos for users from such a large amount of content.

  Video content has a description associated with it, often in text format, which allows search based on keywords through a set of free text fields or labels. Thus, each user who wants to upload a new video can use their search engine or an external search engine (such as “Google”, “Bing”, or “Yahoo!”) You are asked to provide a description that allows other users to find your video.

  However, this solution is not sufficient because the description is completely subjective because it is provided by the user. In particular, they represent individual user interests and may be in various languages, and in a given language may include selection of keywords that include multiple possible synonyms. Thus, to find videos, the user must select the appropriate keywords without the tools to determine them. How can you know the label and terminology selected by the user who uploaded the video?

The resulting problems are the following two.
-Search does not necessarily provide the most relevant video content.
-The video may not be fully visible to the user performing the search.

  The object of the present invention is to improve this situation.

In order to do so, its first purpose is a method for supporting video content retrieval by a user through a communication network, comprising:
-Determining a first content item by the user;
-Automatically extracting semantic data from the first content item;
-Automatically generating a query for at least one service in response to the semantic data to allow retrieval of a set of videos at the at least one service;
-Presenting the set of videos to the user.

  According to one embodiment of the present invention, the method is configured to determine the set of the semantic content between the first content item and the description associated with the video content in the set. Can also include selecting a subset of the video from.

  At least one service can be determined in response to semantic data.

  These steps can be triggered by selection of the first content item (CZ) of the man-machine interface (WB) and by selection of software functions accessible through the man-machine interface.

  The step of automatically extracting semantic data may consist of extracting a “subject-verb-object” relationship. The generated query can be constructed based on the relationship.

  The present invention also aims at a communication terminal comprising processing means for carrying out the method described above, and a computer program comprising coding means for carrying out this method.

The invention comprises means for allowing a user to select a first content item, and automatically extracting semantic data from the first content item;
-Automatically generating a query for at least one service in response to these semantic data, allowing said at least one service to retrieve a set of videos;
-It is also aimed at a man-machine interface for a communication terminal, with means allowing to trigger the step of presenting the set content of the video to the user.

  In this way, based on the first content item, which is generally in text form, the present invention removes the user's subjectivity to improve video content visibility and search relevance, Or at least allow to reduce. The present invention makes it possible to reduce the semantic gap between video content and search objects.

  The invention and its advantages will become more apparent from the following description with reference to the accompanying drawings.

It is a figure which shows the man machine interface which implements this invention. 3 is a flow diagram schematically showing a possible chain of steps of the method according to the invention. FIG. 3 is a diagram illustrating an example of text content and extracted semantic data.

  FIG. 1 shows a web browser that allows a user to access a web page from a URL (Unified Resource Locator). This browser presents the user with a man-machine interface or “window” WB. In this window, a plurality of parts including a browser space WP in which a page (including video contents) is displayed and a menu area MZ are displayed.

  This menu area can itself contain a drop-down menu SM and an icon IM. Through this menu area MZ, the browser provides the user with access to various functions. A given function may also be accessible through a drop-down menu, an icon, or both. This is an ergonomic choice made by browser developers.

  In addition, these and / or other functions may be accessible to the user through a context menu CM. Typically, in Microsoft's Windows (TM) operating system, such a context menu is opened with the right mouse menu button, and its elements depend on the position of the mouse pointer and the state of the browser doing.

  In the example of FIG. 1, text content is displayed in the browser space WP. The man-machine interface WB makes it possible to select all or part of this text content. This selection can be performed by a known method using a mouse or any other pointing device or a keyboard. In the depicted example, only the CZ region is selected.

  When the first content CZ is selected, the user can select a function corresponding to the present invention from the menu of the man-machine interface WB. This function can be selected from a drop-down menu SM, an icon IM, or a context menu CM. This selection may consist of clicking on the icon IM or selecting an element from the drop-down menu SM or the context menu CM.

  The present invention can also be applied to software applications other than browsers. In particular, the present invention can be applied to a file display application. A user reading an Adobe “PDF” formatted e-book (or “e-book”) or file selects a content area as described above and from the menu in the appropriate software application based on the selected content. A function can be triggered.

  The software application may also be a word processor such as Microsoft's “Word” software.

  Similarly, the present invention can be applied to contents other than text contents. It is therefore possible to select an image, sound or video.

  Mechanisms other than selection by the user to determine the initial content may be possible. For example, the determination may be automatic depending on the displayed web page, the video being played, the content of the email being read, etc.

  The step of determining the first content item CZ is shown in step E1 of FIG.

  Selecting the first content item CZ and the appropriate function triggers a first algorithm step that automatically extracts semantic data from the first content item.

  This step is shown as step E2.

  This step can be subdivided into several substeps.

  The first sub-step E2a consists in determining the type of content. That is, whether the content is text, video, image, audio, or the like.

  Depending on the content type, various algorithms can then be used to extract semantic data. These algorithms can be implemented by separate software modules.

  The next sub-step of the extraction step E2 can depend on the content type.

  In the case of text content, the second sub-step E2b may consist of extracting semantic raw data from the content item CZ.

  These semantic raw data can be extracted for each sentence or for each block of sentences.

  The extraction of sentence blocks is coordinated by the presence of personal pronouns that refer to previously identified entities. Since the subject of each sentence can be said to be the same, it may be possible to identify consistent sentence blocks on the same topic.

  Not all words need to be extracted. Some words can be omitted if they do not help to understand the text.

The extraction can be performed in a plurality of steps shown below.
-Deletion of so-called "stop words": coordinate conjunctions, etc.
-Use statistical and natural language algorithms to analyze the remaining information and extract a set of keywords with discriminative meaning. For example, in the description of a person, the word “eye” is not discriminatory. Because everyone has “eyes”. On the other hand, the word “one eye” is discriminatory. To this end, a corpus that can obtain information about the frequency of words in a set of documents can be used. In other words, the lower the frequency, the higher the speculative importance of the word.

  Based on these semantic raw data, refined semantic data can be determined in a third sub-step E2c. These sophisticated semantic data may be “concepts”; names of people, companies, places (city, country, region, etc.); dates, etc.

  A “concept” can be viewed as the main semantic data from a set of semantic raw data. They can be determined in various ways. For example, each semantic raw data can be associated with an occurrence counter across the text content CZ. The concept is then semantic data corresponding to the one with the highest number of occurrences.

  It also includes raw data from synonyms and various semantic levels, especially broader terms (ie, more general semantic raw data that includes one or more more specific semantic data items). It is possible to use an ontology that takes into account

  The concept can then be determined by the number of occurrences of semantic raw data aggregated based on synonyms and broader terms. For example, the concept of “fruit” can be derived from textual content that includes the occurrence of raw data “orange”, “grape”, “banana”, and the like.

  In order to determine the category to which the text content CZ belongs, it is possible to derive the most important concepts, for example those related to the one with the highest number of occurrences. This category can be searched considering a series of broader terms of previously determined concepts.

  Names of people, companies, places, etc. can be determined from the dictionary. This dictionary may be internal to the software means of the present invention or may be remote on the Internet in the form of a web service or other technical device (such as a database).

  FIG. 3 shows the correspondence between the example text content CZ and the extracted semantic data. In this example, the text content CZ is an area selected by the user on the web page displayed in the browser space WP. The figure shows a gray background.

  After triggering the function of the present invention, a “pop up” window PW is displayed, displaying the most relevant video content VC and other relevant video content RVC. The window PW also displays “category”, concept, name of person and place (“state or country”), and semantic raw data (“tag”). This information may not be displayed if it does not make much sense to the user and is more useful internally. However, it is displayed here for clarity.

  The category “business” (trade, trade) is determined in this example based on the extracted concepts. Therefore, the concepts “advertisement” and “sale” have the broader term “business”. This reasoning is further strengthened by the existence of raw data such as “Low”, “Market”, “Bill Brokerage”, “Shilling”.

  Since multiple concepts arise in this semantic field, the method of the present invention can now automatically determine that the text content CZ is business relevant.

  Sub-step E2d consists of weighting the semantic data. Depending on each semantic data and the number of occurrences of its synonyms and broader terms, the importance of the semantic data can be identified and a weight assigned to represent the text content CZ.

  Higher weights can also be assigned.

  In practice, this sub-step can be performed simultaneously with sub-step E2c.

  Also, ranking based on these weights for use in later steps may include semantic data filtering to preserve only those whose weights exceed a certain threshold or limit. it can.

  In one embodiment of the present invention, it is possible to perform grammatical analysis of text content CZ to determine the grammatical functions of semantic data and to construct logical links between them. is there.

  This link can be formalized by a “subject-verb-object” relationship.

  The method of the present invention identifies “subject”, “verb”, and “object” functions within a sentence or group of words and establishes such a link between three extracted elements by semantic analysis. Try to do. Of course, various techniques are possible.

One possible approach may consist of:
-Extracting a subject from a sentence or group of words. This subject, the person who acts, can be identified as an individual, a known organization, an actual location, or the like. A dictionary can be used for such identification.
-Detection of verbs near the extracted subject. Again, a dictionary can be used to identify the form of the verb. It is also possible to consider the use of a given verb and various forms. Depending on this format, supplemental semantic information can be extracted to see if this is related to behavior (action verb) or state (passive, past participle).
-Extract keywords and concepts near the verb to identify the "purpose" of the action.

  To improve the efficiency of this process, one can plan to consider only transitive verbs. As a result, a more direct “subject-verb-object” relationship can be constructed. To do so, the dictionary used to extract verb forms may contain only transitive verbs.

  It can also be planned to make the creation of the “subject-verb-object” relationship subordinate to the determination of the object belonging to the corpus belonging to the dictionary. Following the “subject” example, this “object” may be an individual, a known organization, an actual location, or the like.

  Next, step E3 may consist of determining at least one service based on semantic data.

  In fact, some video content offering services are specialized in one theme (sports, news, artistic short films, education, etc.) and rely on semantic data extracted from text content CZ Thus, it may be appropriate to target the most relevant services.

  Additionally, if there are no specialized services in the systematically or extracted topics or if they are insufficient, general services such as DailyMotion or Youtube can also be searched.

  Alternatively, services can be statically identified by configuration.

  Step E4 consists of automatically generating a query for the identified service (s) based on semantic data.

  Typically, this can include constructing an HTTP (Hypertext Transfer Protocol) query that includes all or part of the semantic data extracted as parameters.

  These data can be used to construct a single query or multiple queries. Thus, it is possible to construct as many queries as semantic data with weights exceeding a certain threshold.

  Also, in an implementation configured by extracting a “subject-verb-object” relationship, it may be possible to construct different queries for each relationship.

  All of the text content CZ can be browsed by sentence or word block. For each sentence or word block, if there is a “subject-verb-purpose” relationship, they can be used to construct a query. Otherwise, the query can consist of raw semantic data.

  These queries allow to retrieve a set of videos that are available for the identified service (s).

  These videos are then presented to the user at the man-machine interface WB, more specifically in the browser space WP.

  In one embodiment of the present invention, a semantic proximity criterion between text content CZ and the retrieved video can be implemented.

  Semantic proximity can be calculated based on the semantic data extracted from the text content CZ in the previous step and the semantic content associated with the retrieved video. These data may be raw data saved by the user, such as a “tag”, “title” field, “description” field, and the like.

  It is also possible to perform an analysis of the extracted video.

  Thus, the texture (sea, sand, ice, etc.) and / or color can be determined fairly easily.

  It is also possible to implement a shape recognition algorithm. They can attempt to determine a known format from fixed information (ie, a portion of the shape that does not change with viewing angle).

Various techniques can be used together to obtain a synergistic effect.
-For example, determining a seal by shape analysis can make it possible to confirm that the detected texture is actually ice. Second, errors can be reduced and the reliability of the results can be increased.
In another example, adding color may be possible by correlating color determination with shape determination. Therefore, it is possible to determine the existence of a “black horse”, which is a discriminatory concept rather than a simple horse determination.

  It is possible to perform an exhaustive analysis of the image from the video (or successive images). Thus, the level of ambient light can be analyzed and semantic information can be extracted therefrom. For example, if it is detected that the scene is outdoors, it can be determined whether it is daytime or nighttime.

  By comparing this information, it is possible to determine the distance, especially by using ontology.

  Also, various methods are possible for determining such distance.

In particular,
-A heuristic metric between given ontological concepts based on the distance treated in the graph. This distance may be the number of nodes on the graph where the nodes are keywords or concepts. Certain pairs are directly connected by links. The link may be directional or symmetric. When the link between two words is symmetric, the relationship between the two words is stronger. This distance can make it possible, for example, to quantify the similarity of two concepts. This can also serve the purpose of disambiguating.
The TF-IDF (word appearance frequency-reverse document frequency) method is a weighting method often used in information retrieval, particularly text comparison. This statistical measurement makes it possible to evaluate the importance of the words contained in the document to the set or corpus. The weight increases in proportion to the number of times the word appears in the document. It also changes based on the word frequency in the corpus. To evaluate the relevance of documents with respect to user search criteria, search engines often use original formula alternatives.

  For example, if the incidence of the word “dinosaur” in the IDF corpus is low, this may mean that it is more discriminating and more important than the more general word, and this word is If it appears in each of the “TF” lists (a list linked to a corpus extracted from context and a list linked to a corpus extracted from a video description) Lower weight can be given. The fact that words are found in both lists makes it possible to define stronger or weaker links based on IDF values.

  Also, the two methods can be combined. For example, in the first list of words extracted from the context, the word “dinosaur” is found, which has a relatively high weight in the IDF corpus. In the second list extracted from the video description, the word "tyrannosaurus" is found. The graph shows that there is a symmetric link at the two hops (number of nodes in the graph). Thus, the context of this video can be thought of as showing similarity to the IDF value / number of hops.

  Therefore, the overall similarity can be calculated by applying this algorithm to each word.

  Based on the distance of each retrieved video, they can be ranked to present them to the user in the most reasonable way. Thus, the video with the closest proximity to the text content CZ is presented before the other related videos that are retrieved.

  It is also possible to select a subset of the retrieved video based on proximity criteria.

  This can include selecting only a predetermined number of videos in response to a ranking based on their semantic distance from the text content CZ.

  It can also include selecting only videos that correspond to semantic distances less than a predetermined threshold. Thus, if none of the retrieved videos are close enough, this embodiment may not show the video content to the user.

  Other alternatives and embodiments are possible and available to those skilled in the art. Therefore, it cannot be understood that the present invention is limited to the implementation described above.

Claims (9)

A method for supporting a user to search for video content through a communication network,
-Determining a first content item (CZ) by said user;
-Automatically extracting semantic data from the first content item;
-Automatically generating a query for at least one service in response to the semantic data to allow the at least one service to retrieve a set of videos;
-Presenting the set of videos to the user.   2. The method further comprising selecting a subset of videos from the set based on a semantic proximity criterion between the first content and descriptions associated with the videos in the set. The method described in 1.   The method of claim 2, wherein the at least one service is determined based on the semantic data.   The step can be triggered by selection of a first content item (CZ) of a man machine interface (WB) and by selection of software functions accessible through the man machine interface. The method of any one of these.   5. A method according to any one of claims 1 to 4, wherein the step of automatically extracting semantic data comprises the step of extracting a "subject-verb-object" relationship.   The method of claim 5, wherein the generated query is constructed based on the relationship.   The communication terminal provided with the processing means for implementing the method of any one of Claims 1 thru | or 6.   The computer program provided with the processing means for implementing the method of any one of Claims 1 thru | or 7. Means for allowing a user to select a first content item (CZ); and- automatically extracting semantic data from the first content item;
-Automatically generating a query for at least one service in response to the semantic data to allow retrieval of a set of videos at the at least one service;
A man-machine interface for a communication terminal comprising means for enabling the user to trigger the step of presenting the set of videos to the user.

Images